Provenance

Definition

Provenance is the traceable record of where a judgment came from: who made it, under which task specification and guideline version, at what time, and what happened to it afterwards. In machine learning it applies at two levels — the provenance of the material a model was trained on, and the provenance of each individual human judgment inside a training or evaluation set. The second is what makes a quality claim checkable. Without it an aggregate score is an assertion; with it, any number can be traced back to the specific records that produced it.

A usable record is append-only and captures identity, task, version and time together. Identity means a verified contributor rather than a display name. Task means the specification and the guideline revision in force when the judgment was made, since guidelines change and old judgments were made under old rules. Time means an ordering that makes review and dispute resolution possible. Where the record also covers the review chain — who checked the work, who arbitrated a disagreement, what changed — it supports the separation of duties that lets a measurement mean something, because the party reporting quality is then demonstrably not the party that produced the work.

Provenance is also what makes a result reproducible by someone else. If the underlying judgment records travel with the statistics computed from them, a recipient can recompute the numbers instead of trusting them, and can inspect the specific items behind a disputed figure. Procurement and regulatory pressure has pushed in the same direction: buyers increasingly ask where training material and human judgments originated, and a record answers that in a way a recollection cannot.

How DeepenSkill approaches it

Every vendor says their experts are good; the record is what lets you check who measured. Everyone clears the same bar before starting: credentials verified, a standardized domain test passed, then a practical task scored against gold reference data. On live work the team measuring the experts is separate from the team running them — the platform rejects a validation report authored by the party that delivered the work — agreement statistics are reported with the method named, the item and rater counts stated and a 95% interval attached, and every judgment traces to a verified contributor, a task spec and a timestamp. See expert human-in-the-loop work, the annotation work it records, and a sample verification bundle you can recompute yourself.

FAQ

• What is data provenance?

  The traceable record of where a judgment came from: who made it, under which task specification and guideline version, at what time, and what happened to it afterwards.

• What should a provenance record contain?

  Identity as a verified contributor rather than a display name, the task specification and guideline revision in force at the time, a timestamp, and ideally the review chain — who checked the work and who arbitrated any disagreement.

• Why does provenance matter for AI training data?

  Because it turns a quality score into something checkable. If the underlying judgment records travel with the statistics, a recipient can recompute the numbers instead of trusting them.